Hesitant neural gas for supervised and semi-supervised classification

Abstract

Neural Gas is a neural network algorithm for vector quantization. It has not arbitrary established network topology, instead its topology is changing dynamically during training process. Originally, the Neural Gas is an unsupervised algorithm. However, there are several extensions that enables Neural Gas to use the information about sample's class. This significantly improves the accuracy of obtained clusters. Therefore, the Neural Gas was successfully used in classification problems. In this paper we present a novel method to learn the Neural Gas with fully and partially labelled data sets. Proposed method simulates the neuron's hesitation between membership to the classes during the learning. Hesitation process is based on neuron's class membership probability and Metropolis-Hastings algorithm. The proposed method was compared with state-of-art extensions of Neural Gas on supervised and semi-supervised classification tasks on benchmark data sets. Experimental results yield better or the same classification accuracy on both types of supervision. © 2013 Springer-Verlag.